You study whether interventions work.
We make the outcome measure auditable.
How it works
Every study follows the same pipeline — from design to independently verifiable results.
Power calculator, pre-registration, primary outcome freeze. Server-enforced — no post-hoc switching.
Scaffold-aware proficiency tracking. Knows the difference between an unaided answer and a helped one.
Hash-chained records, standalone verifier, deterministic replay. Any third party can audit.
Capabilities
Proficiency measurement that distinguishes helped from unaided performance. Certifies mastery only when evidence survives scrutiny.
Item calibration with verifiable certificates. Fit diagnostics with root-cause analysis, not folk cutoffs.
Known-truth validation. Runs your engine against synthetic learners alongside baselines — reports where it wins and where it loses.
Content validation against domain rules. Every rule teaches: names the failing element and says how to fix it.
Validated
Synthetic data calibrated to published parameters. The platform recovers the known effect in every case.
VanLehn 2011
d = 1.003
Mueller & Dweck
d = 2.20
Hattie 2009
ρ = 0.83
Chi 2009
27.4%
Kraft 2021
N valid
Verified on real data
Synthetic recoveries prove the engine finds known truths. These calibrations go further: real public data, novel numbers, and a certificate that lets you rerun everything yourself.
ASSISTments skill builders, 2012–13 public release
ASSISTments skill builders, 2012–13 public release
Research
Open research behind the platform. All published under open licenses.
Why open education AI needs one more layer — and what it's made of.
Kumar Srivastava · July 2026 · CC-BY-4.0 · Coming soon
Scaffold-conditioned proficiency estimation with verifiable evidence trails.
EdArXiv Preprint · 2026 · Coming soon
Architecture, validation, and the case for separating measurement from instruction.
EdArXiv Preprint · 2026 · Coming soon
Free for research use. No dark patterns. Pseudonymous by default.